| Massive remote sensing data requires efficient machine interpretation methods.The segmentation accuracy of remote sensing images directly affects the interpretation level of remote sensing images.High-resolution remote sensing images have become one of the important research directions of remote sensing applications due to their rich ground detail information.The current image segmentation methods usually adopt supervised learning,that is,the machine needs to learn the labeled samples before it performs automatic annotating.For the actual needs of large-scale sample labeling,"crowdsourcing" labeling with the help of crowd workers is gradually becoming mainstream.However,the labeled samples derived from this labeling approach often contain a considerable proportion of incorrect labels,i.e.,label noise.For high-resolution remote sensing images,due to the diversity of surface features,the complexity of surface detail information,and the professionalism of sample labelers,the mislabeling problem is more likely to occur.Therefore,for high-resolution remote sensing images,how to perform accurate segmentation with noisy annotations has become an unavoidable problem in current theoretical researches and practical applications.For this reason,this dissertation focuses on high-resolution remote sensing images and carries out image segmentation research based on learning with noisy labels.The main works and innovations of this dissertation are as follows:(1)Before we study the learning of the image samples with noisy annotations,a remote sensing image segmentation model with good performance is indispensable.Aiming at addressing the problems of insufficient training and structural stereotype,which are rooted in the encoder-decoder-based image segmentation model,this dissertation proposes an improved high-performance remote sensing image segmentation model.The topological structure of the residual network is introduced to all convolutional layers in the encoderdecoder model to alleviate the problem of insufficient training caused by vanishing gradients.Besides,to solve the problem of structural stereotype,a strategy of random position sampling in training is proposed to make the sample training fair,and an ensemble inference method in inference is proposed to solve the jaggedness in the inference label map.The method yields good performance on two datasets in the ISPRS remote sensing image semantic segmentation competition.(2)For the samples of label noise,this dissertation researches the learnability of the main pattern(interest pattern).The research involves how to learn for the main pattern when multiple patterns coexist.Based on the optimization of gradient descent,this dissertation reveals two important factors that affect samples learning,i.e.,regularity and scale.This study shows that the deep neural networks tend to learn large-scale regularity patterns first,then small-scale regular patterns,and finally irregular patterns.This finding provides support for the construction of the learnability theory for noisy samples.Next,this dissertation gives and proves the learnability conditions of the main pattern in the training environment with noisy labels,that is,the noise pollution rate should meet what kind of conditions under what type of noise.(3)In the learning environment with noisy labels,to solve the problem that traditional overfitting methods tend to fail due to lack of clean verification set in practical tasks,this dissertation proposes a method,which is called limited gradient descent(LGD),that does not require a validation set to prevent overfitting.This method artificially prepares a few of almost purely false samples as a reverse pattern.Using the previous research conclusion,that different patterns have different learning progress,we can estimate the optimal generalization timing of the main pattern by monitoring the learning progress of the reverse and the remaining samples.Therefore,this LGD method can automatically decide when to stop training without involving a clean validation set.If the learnability condition of noisy samples is met,one can choose to use the LGD method for learning.At the same time,according to this condition,we also provide the upper bound of the ratio of the selected reverse samples,which is a theoretical guarantee for the application of LGD.Experimental results show that the LGD learning method avoids involving a validation set and has robust performance in preventing overfitting.(4)For the learning of remote sensing image samples with label noise,this dissertation proposes a noise-robust objective function.A good noise-robust objective function can effectively delay the overfitting process of noise and promote the generalization of the main pattern.The proposed objective function achieves this goal from two aspects: on the one hand,it uses a robust loss function for noise,which can ensure faster training speed while suppressing noise;on the other hand,it implements a regularization approach of linearly augmenting samples to suppress noise.The experimental results show that the objective function constructed in this dissertation can significantly increase the optimal generalization level of the main pattern.(5)Integrating the above research results,this dissertation embeds a high-performance image segmentation model and a noise-robust objective function into the LGD learning framework to construct a high-resolution remote sensing image segmentation method based on the learning with noise labels.The experimental results show that the proposed method has strong generalization performance and does not need to employ a clean validation set to prevent overfitting.From the perspective of compatibility,the method proposed in this dissertation can adapt to the learning of both noisy-label and clean samples.To sum up,this dissertation focuses on supervised learning for high-resolution remote sensing image segmentation tasks and carries out the research on learning theories and methods w.r.t.noisy-label samples.Aimed at the problems of insufficient training and structural stereotype in current deep learning models,overfitting problems and highresolution remote sensing image segmentation methods with noisy training environments,we conduct an in-depth theoretical analysis and effective solutions.This provides strong theoretical support and a good choice of methods for improving the segmentation accuracy of high-resolution remote sensing images and improving their application level. |
